Skip to main content
DoneThat

AI Adoption GuideHRReward

Personalized total-rewards statement

LLM generates per-employee narrative covering base pay, equity, benefits, and growth in pay over time.

HR processPlanSourceSelectHireOnboardDevelopRewardExit

By Don, DoneThat’s AI coach · updated

The statement is a cited pack, not a story

A total-rewards statement is ready for an employee only when three claims, if they appear at all, point at named records. Base pay cites the pay file. Equity cites a grant the employee actually holds. Benefits cite the election on file for the statement period. Growth in pay over time, if you include it, cites dated amounts from that same pay history. It does not cite a model estimate.

Quality here means the pack is source-true. Compensation still approves every pack. Communications still sends. The model does not auto-send. A fluent paragraph with no cite is a defect, even if it matches last year's comms template.

Keep range position off the employee PDF unless Compensation has already approved a cited figure for that person. Range work belongs with live market benchmarking. If a sentence cannot point at the pay file, a grant, or an election, it does not belong on the statement.

What you load: pay file, grant, and election

Load records first. Then generate. Do not paste a name into a prompt and ask for a personalized total-rewards story. The model may only narrate fields that arrived in the extract.

Pay. Pull current base, and any documented cash allowances you intend to show, from the HRIS pay file. Keep employee ID, amount, currency, and effective date on the cite. Workday is one common system of record for that file. The cite is the pay-file row, not a manager recollection and not an all-hands slide.

Equity. Pull grant ID, instrument type, the share or unit count as stored, and vest dates from equity administration. Carta is one common system of record for grants. If the employee has no grant in the file, the equity block stays empty. Do not describe a standard grant for the level.

Benefits. Pull plan names and coverage tiers from the benefits election for the statement period. Fidelity is one common administrator. Some elections live in the HRIS benefits module instead. If medical was waived, do not invent a plan name. If dental is missing from the extract, leave dental empty.

Performance notes can sit in the working folder because managers will ask. Lattice is one common system of record for reviews. A rating is not a pay cite, not a grant cite, and not an election cite. It does not authorize a growth percent and it does not belong in the employee-facing money blocks.

Label every extract: source system, as-of timestamp, employee ID, and which statement period it covers. If two systems disagree, do not average them in prose. Stop and let Compensation pick the record of truth before you draft.

Offer-stage recommendations are a different artifact. Do not reuse copy from comp recommendation per offer unless every number still matches the current pay file and grant file. New hires who need a walkthrough after send can use a new-hire conversational assistant that reads the approved pack. That assistant must not read a newer unapproved draft.

Draft, blank, approve, then send

Run the pack in this order every cycle.

First, load pay, equity, and benefits for the named employee and the statement period. Put extract time and source system on the pack cover so reviewers can see staleness.

Second, draft the narrative so that each money, grant, and election sentence carries its cite: employee ID, record type, identifier, and as-of date. The citation can be a footnote, a hidden metadata block, or an appendix. What matters is that Compensation can open the source and see the same number.

Third, leave a line empty when its cite is missing. Do not replace the blank with a hedge ("your equity is an important part of your package") and do not gray-box a guess. Empty is the honest state. The rest of the pack can still move.

Fourth, send the full pack to Compensation: filled lines, empty lines, and the cite list. Empty lines are in scope for approval.

Fifth, Communications sends only after that approval, through the agreed channel, to the agreed audience. The draft is not sent. Nothing generates-and-mails in one job.

Three failure modes show up in real reviews.

A line with no pay-file cite. "Your base compensation reflects your impact this year" has no amount, no effective date, and no pay-file pointer. Strip it. If the pay file did not load, the base-pay line stays empty and the pack still goes to Compensation. Do not hold the whole wave for one missing file unless Compensation asks you to. Do not fill base from a planning spreadsheet just this once.

Treating the draft as sent. A generated file in a shared drive, a preview to the employee, a manager paste in chat, and a calendar hold that attaches the PDF all count as send. Until Compensation has approved and Communications has executed send, the artifact is a draft. Turn auto-send off.

Inventing a pay-growth percent. If the file stores two dated bases, you may state those two amounts and dates. You may not mint a percent, even when the arithmetic looks obvious. If Compensation does not already store an approved growth percent, that line stays empty.

When growth in pay has no stored percent

Employees often want a single figure for how much pay moved. The quality rule does not bend for that preference. A percent looks like a fact. An unsourced percent is still invented.

Allowed: state base on date A as amount A, cited to the pay file, and base on date B as amount B, cited to the pay file.

Not allowed: "your base grew X percent year over year" when that percent is not a stored, Compensation-approved field. Not allowed: "approximately," "about," or a rounded figure that someone computed in the draft.

If history is incomplete, one year missing, a mid-year entity or currency change without a documented method, leave the growth block empty rather than interpolating. Compensation can add a percent later by writing it into the file and re-approving.

Pay-equity findings are also out of scope for this statement. Detection and remediation tracking belong on continuous pay-equity monitoring. The statement reports what the employee is paid according to the file. It does not quietly close a gap, imply a future adjustment that is not already in the pay file, or explain a peer's pay.

One analyst's mid-year packet

This is an illustrative sequence, not a measured case.

A senior analyst, Maya, is in the mid-year comms wave. HR loads three extracts. The pay file has current base, currency, and effective date, plus prior-year base with its effective date. Equity administration returns one RSU grant with a grant ID and vest dates. Benefits returns medical and retirement elections for the plan year. There is no Compensation-approved growth-percent field in the pay file. A review rating exists in the performance system.

The draft should state current base with the pay-file cite, name the RSU grant by ID with the equity cite, and name the elected medical and retirement plans with the election cite. It may list the two dated bases. The growth-percent line stays empty. The rating does not appear as if it explained the pay number.

Suppose the equity extract fails that morning. The grant paragraph stays empty. Pay and benefits can remain. Compensation still reviews the pack, including the empty equity block. Communications does not send a version that fills in a share count from a level guideline or from last year's statement.

Suppose a well-meaning editor adds a year-over-year percent because the two bases are higher this year. That sentence is a defect. Remove it. Do not round. Empty stays empty until Compensation stores and approves a percent in the file.

After approval, Communications sends. Maya's manager can receive the same approved pack. Nobody forwards yesterday's draft.

Workday, Fidelity, Carta, and Lattice are examples of where these records often live. The quality rule does not change with the logo: cite the pay file, the grant, and the election, or leave the line blank. Compensation approves the pack. Communications sends. Do not auto-send. Do not invent a pay-growth percent.

Is this worth automating for you?

Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.

DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.

Measure the baseline first